On Wed, Sep 7, 2016 at 1:47 AM, Robert Haas <robertmhaas@gmail.com> wrote:
> On Mon, Sep 5, 2016 at 8:57 PM, Masahiko Sawada <sawada.mshk@gmail.com> wrote:
>>> What performance difference does this make, in a realistic use case?
>>
>> I have yet to measure performance effect but it would be effect for
>> very large frozen table.
>
> I think if you are proposing this patch because you think that the
> existing code won't perform well, you definitely need to submit some
> performance results showing that your approach is better. If you
> can't show that your approach is practically better (rather than just
> theoretically better), then I'm not sure we should change anything.
> It's very easy to screw up the code in this area and we do not want to
> risk corrupting data for the sake of an optimization that isn't really
> needed in the first place.
>
> Of course, if you can prove that the change has a significant benefit,
> then it's definitely worth considering.
>
I understood, thank you.
I've measured the performance benefit of this patch by following steps.
1. Create very large table and set all-visible flag to all blocks.
2. Measure the execution time of vacuum that skips to scan all heap pages.
* 1TB Table(visibility map size is 32MB)
HEAD : 11012.274 ms (00:11.012)
Patched : 6742.481 ms (00:06.742)
* 8TB Table(visibility map size is 64MB)
HEAD : 69886.605 ms (01:09.887)
Patched : 35562.131 ms (00:35.562)
* 32TB Table(visibility map size is 258MB)
HEAD: 265901.096 ms (04:25.901)
Patched: 131779.082 ms (02:11.779)
Since current HEAD could scan visibility map twice, the execution time
of Patched is approximately half of HEAD.
But when table is several hundreds gigabyte, performance benefit would
not be large.
Regards,
--
Masahiko Sawada
NIPPON TELEGRAPH AND TELEPHONE CORPORATION
NTT Open Source Software Center